fix(compress): reserve output tokens in the compaction threshold (#23767, #43547)

[kshitijk4poor]

Summary

The compaction trigger compared estimated input against context_length × threshold, but the provider reserves max_tokens of output out of that same window. With a large max_tokens (e.g. 65536 on a custom provider), the usable input budget is materially smaller than the raw window — so a session could hit a provider 400 ("context length exceeded") before compaction ever fired. The threshold is now based on the effective input budget (context_length − max_tokens). (Mode B of #23767; fixes #43547.)

Salvage of #43651 by @kyssta-exe — reimplemented on the current threshold surface (see below).

Why a fresh commit (partly-superseded)

The original PR edited an inline int(context_length * threshold_percent) in update_model/__init__. That code was since refactored into the static _compute_threshold_tokens() helper (for the small-window 85% guard, #14690), and the threshold reset now runs in update_model alongside the #50137 calibration reset. A cherry-pick would conflict and reintroduce the old shape, so this re-implements the one surviving design point (output-token reservation) on the current surface.

Changes

• agent/context_compressor.py: _compute_threshold_tokens() gains an optional max_tokens param and subtracts it from the effective window before both the percentage and the #14690 small-window 85% guard; self.max_tokens stored in __init__ and reused by update_model (optional explicit override). max_tokens=None (provider default) → no reservation → full-window behavior, byte-identical to before.
• agent/agent_init.py: pass max_tokens=agent.max_tokens at construction.
• tests/agent/test_context_compressor.py: 3 tests (reservation lowers threshold; small-window floor composition; max_tokens ≥ context_length falls back to full window).

Design note vs original PR

The original threaded max_tokens as a required update_model kwarg but updated only the one construction caller — the other 6 update_model callers would have zeroed the reservation on every switch. This version stores max_tokens and reuses it across switches (the output cap is a user setting, not model-specific), so the reservation survives /model switches; an explicit kwarg can still override it.

Validation

	Result
threshold/max_tokens/compress tests	117 passed
existing threshold tests (None path)	unchanged ✓
ruff (diff vs main)	clean
Negative check	3 new tests fail on main without the fix ✓
E2E (real imports)	200K+65536 → threshold 67232 (not 100000); None → 100000 unchanged; switch preserves reservation; explicit override works

Interactions verified orthogonal: #50137 (calibration reset in update_model) and #50136 (tool-output persistence cap, different layer).

Part of #23767 — this is the last of the 6 failure modes; #23767 can close once this lands.

Credit

• @kyssta-exe — fix(compress): account for output-token reservation in compaction threshold (#43547) #43651 originally proposed the output-token reservation in the compaction threshold. The surrounding threshold code was refactored since (BUG: Context auto-compression never triggers when context_length == MINIMUM_CONTEXT_LENGTH (64000) #14690 small-window guard), so this re-implements that design point on the current _compute_threshold_tokens helper.

Closes #43547.

Infographic

Image generation is unavailable in this environment (FAL_KEY unset, no managed-provider credits); to be attached once available.

[github-actions]

🔎 Lint report: salvage/output-token-reservation-threshold vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 11356 on HEAD, 11354 on base (🆕 +2)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 5969 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.